Migration of Graphics Processing Unit (gpu) States

1. A method for live migration of graphics processing unit (GPU) states, the method comprising: receiving a snapshot of a GPU state of a source GPU; restoring the snapshot of the source GPU at a destination GPU; receiving a predicted downtime comprising a time that a number of additional recorded GPU commands to be received from the source GPU is predicted to be below a threshold level based on a predicted behavior of an application or a game, wherein the predicted behavior is based on a previous behavior of the application or the game; receiving GPU commands recorded after the snapshot and up to and including the predicted downtime from a relay at the destination GPU; replaying the received recorded GPU commands at the destination GPU up to the predicted downtime; and switching the relay to connect the destination GPU to a client at the predicted downtime.

2. The method of claim 1 , further comprising receiving a map linking GPU resource identifiers to logical resource identifiers, each of the GPU commands comprising an object with a GPU resource identifier that is to be mapped to a particular logical resource identifier.

3. The method of claim 2 , further comprising utilizing the map to order GPU commands in a thread for execution, each thread containing one or more logical resource identifiers.

4. The method of claim 2 , wherein GPU commands with logical resource identifiers that are lower are given higher priority over GPU commands with logical resource identifiers that are higher in value.

5. The method of claim 1 , the destination GPU located in a machine that is closer in proximity to the client than the source GPU.

6. The method of claim 1 , further comprising processing GPU commands from a plurality of clients at the destination GPU.

7. The method of claim 6 , further comprising dynamically allocating resources by detecting an underutilized GPU and migrating the client to the underutilized GPU.

8. The method of claim 6 , further comprising dynamically allocating resources between a software renderer and a hardware renderer by migrating the client between the software renderer and the hardware renderer.

9. The method of claim 6 , further comprising dynamically allocating resources by processing GPU commands from a plurality of clients at the destination GPU.